Best local LLM for a laptop with only 8GB RAM
If you only have 8GB of RAM, your choice of local model matters more than almost anything else in your setup. Eight gigabytes sounds like plenty until you remember that your operating system, browser tabs, and background apps are already using two to four of those gigabytes before you even open a model. That leaves a tight window, and picking the wrong model means swapping, freezing, or a model that simply will not load.
This guide breaks down which models actually work on 8GB RAM, why that limit exists, and which free tools run them best.
Quick answer: the best local LLM for 8GB RAM
For most people on 8GB RAM, Qwen3 4B (Q4_K_M quantization) is the best overall pick. It is fast enough for real conversations, strong at reasoning and instruction following for its size, and its memory footprint leaves enough headroom for your system to keep running normally.
If you want something even lighter, Phi-4 Mini (3.8B) is the better choice. It has a smaller footprint, an MIT license, and it was built by Microsoft specifically to run well on modest hardware without falling apart on basic tasks.
If you care more about multilingual support or general world knowledge than raw speed, Gemma 3 4B from Google is a strong alternative, though it needs slightly more memory at the same quantization level.
| Model | Parameters | RAM at Q4_K_M | Best for |
|---|---|---|---|
| Qwen3 4B | 4B | ~3 to 4GB | Best all-around pick, long context, reasoning mode |
| Phi-4 Mini | 3.8B | ~3 to 4GB | Lightest footprint, MIT license, fastest responses |
| Gemma 3 4B | 4B | ~4GB | Multilingual tasks, general knowledge |
| Qwen3 1.7B | 1.7B | ~1.5GB | Backup option if 8GB is shared with other heavy apps |
Why 8GB RAM is the floor, not the comfort zone
Industry guidance converges on the same rule of thumb: 8GB RAM is the practical minimum for local LLMs, and it limits you to the 3B to 4B parameter range at 4-bit quantization. Below that range, models get noticeably worse at following instructions or holding a coherent conversation. Above it, you risk the operating system swapping memory to disk, which turns a usable chat experience into a slideshow.
The key number to understand is quantization. A model’s parameter count tells you how large it is in theory, but quantization decides how much RAM it actually uses. Q4_K_M, the most common quantization format for everyday use, shrinks a model by roughly 75 percent compared to its full precision version while keeping most of its quality intact. A 4B parameter model at Q4_K_M typically needs around 3 to 4GB of RAM just for the weights, with the rest of your 8GB covering context, the operating system, and whatever chat application you are running.
Comparing the best local LLM options for 8GB RAM
Qwen3 4B: the strongest all-around choice
Qwen3 4B currently sits at the top of most recommendation lists for 8GB machines. It outperforms similarly sized competitors like Gemma 3 4B and Llama 3.2 3B on several benchmarks despite being roughly the same size, and it supports a genuinely long context window, which matters if you want to paste in documents or have extended conversations. Qwen3 also ships a thinking mode that trades some speed for better reasoning on harder questions, which you can toggle depending on the task.
Phi-4 Mini: the lightest and most efficient option
Phi-4 Mini trades a bit of raw capability for efficiency and licensing freedom. At 3.8B parameters it is slightly smaller than Qwen3 4B, which means faster responses and a smaller memory footprint on the same hardware. Microsoft built it with resource-constrained devices in mind, and it holds up well for everyday tasks like drafting text, summarizing, and answering general questions. Its MIT license also makes it the simplest option if you ever want to use outputs commercially without worrying about terms of use.
Gemma 3 4B: best for multilingual and general knowledge
Gemma 3 4B is Google’s entry in this size class, and it is a genuinely capable model, particularly for multilingual tasks and general knowledge. Its memory requirements at a given quantization level are close to Qwen3’s, so the choice between the two often comes down to personal preference and which model’s writing style you like better. Gemma’s license terms are slightly more restrictive than MIT or Apache 2.0, so check them if you plan commercial use.
Smaller models, if you need more headroom
If you want to go smaller, models like Qwen3 1.7B or Gemma 3 1B will run with room to spare on 8GB, but you will notice the drop in quality on anything beyond basic chat or simple summarization. They are a reasonable choice only if your laptop is also doing other demanding work at the same time.
Ollama vs LM Studio: which tool should you use on 8GB RAM
The model is only half the equation. You also need software to load and run it, and on 8GB RAM the choice of tool affects how much memory is left for the model itself.
Ollama is the most popular starting point. It runs from the command line, handles model downloads and quantization automatically, and its background service has a relatively small memory overhead when idle, which matters when every gigabyte counts. It is also the easiest option if you eventually want to connect a local model to another app through its API.
LM Studio offers a graphical interface instead of a command line, which makes it friendlier if you are not comfortable with a terminal. It can load any GGUF model from Hugging Face directly, and on Apple Silicon Macs it supports Apple’s MLX engine, which tends to run noticeably faster than the standard approach used elsewhere. The tradeoff is that its desktop app carries more idle memory overhead than Ollama’s background service, which is worth factoring in on an 8GB machine.
Either tool will get a 4B model running comfortably on 8GB RAM. If you want the simplest possible setup and do not mind a command line, start with Ollama. If you would rather click through a visual interface and browse models before downloading, start with LM Studio.
Tips to get the best performance from a local LLM on 8GB RAM
A few adjustments make a real difference on constrained hardware.
- Close unnecessary apps first. Each open browser tab alone can use several hundred megabytes, and that adds up fast on an 8GB system.
- Stick to Q4_K_M quantization. Higher precision formats like Q5 or Q8 require meaningfully more RAM for a quality improvement that is hard to notice in everyday use.
- Keep your context window modest. A longer context consumes additional memory on top of the model weights, so there is no need to max it out for a short conversation.
- Step down in size if a model feels sluggish. If responses slow down or the model starts ignoring instructions, that is usually a sign to try a smaller parameter count rather than fight your hardware.
Frequently asked questions
Can you run a local LLM on 8GB RAM at all? Yes. A 3B to 4B parameter model at Q4_K_M quantization runs comfortably on 8GB RAM, typically at 10 to 25 tokens per second on CPU alone, and faster on Apple Silicon or a dedicated GPU.
What is the best local LLM for 8GB RAM without a GPU? Qwen3 4B and Phi-4 Mini both run acceptably on CPU alone. Expect slower generation than with a GPU, but both remain usable for chat, summarizing, and drafting tasks.
Is 16GB RAM much better than 8GB for local LLMs? Yes, noticeably. 16GB opens up the 7B to 8B parameter range, which handles coding and more complex reasoning with less effort. 8GB is the floor that works, not the comfort zone.
Do I need Ollama or LM Studio, or can I use both? You can use both. Many people start with LM Studio to explore and compare models visually, then switch to Ollama once they want to connect a model to another app or script through its API.
Latest blog: How to Reduce Your AI Coding Token Costs
Bottom line
On 8GB RAM, Qwen3 4B is the strongest all-around choice, Phi-4 Mini is the best option if memory is extremely tight or licensing simplicity matters, and Gemma 3 4B is a solid alternative if multilingual quality is a priority. Pair any of them with Ollama for simplicity or LM Studio for a visual experience, run at Q4_K_M quantization, and an 8GB laptop will handle real, everyday local AI work without needing new hardware.
